1// SPDX-License-Identifier: (GPL-2.0 OR BSD-3-Clause)
2/*
3 * Copyright(c) 2020 Intel Corporation.
4 *
5 */
6
7/*
8 * This file contains HFI1 support for IPOIB SDMA functionality
9 */
10
11#include <linux/log2.h>
12#include <linux/circ_buf.h>
13
14#include "sdma.h"
15#include "verbs.h"
16#include "trace_ibhdrs.h"
17#include "ipoib.h"
18#include "trace_tx.h"
19
20/* Add a convenience helper */
21#define CIRC_ADD(val, add, size) (((val) + (add)) & ((size) - 1))
22#define CIRC_NEXT(val, size) CIRC_ADD(val, 1, size)
23#define CIRC_PREV(val, size) CIRC_ADD(val, -1, size)
24
25/**
26 * struct ipoib_txreq - IPOIB transmit descriptor
27 * @txreq: sdma transmit request
28 * @sdma_hdr: 9b ib headers
29 * @sdma_status: status returned by sdma engine
30 * @priv: ipoib netdev private data
31 * @txq: txq on which skb was output
32 * @skb: skb to send
33 */
34struct ipoib_txreq {
35	struct sdma_txreq           txreq;
36	struct hfi1_sdma_header     sdma_hdr;
37	int                         sdma_status;
38	struct hfi1_ipoib_dev_priv *priv;
39	struct hfi1_ipoib_txq      *txq;
40	struct sk_buff             *skb;
41};
42
43struct ipoib_txparms {
44	struct hfi1_devdata        *dd;
45	struct rdma_ah_attr        *ah_attr;
46	struct hfi1_ibport         *ibp;
47	struct hfi1_ipoib_txq      *txq;
48	union hfi1_ipoib_flow       flow;
49	u32                         dqpn;
50	u8                          hdr_dwords;
51	u8                          entropy;
52};
53
54static u64 hfi1_ipoib_txreqs(const u64 sent, const u64 completed)
55{
56	return sent - completed;
57}
58
59static u64 hfi1_ipoib_used(struct hfi1_ipoib_txq *txq)
60{
61	return hfi1_ipoib_txreqs(txq->sent_txreqs,
62				 atomic64_read(&txq->complete_txreqs));
63}
64
65static void hfi1_ipoib_stop_txq(struct hfi1_ipoib_txq *txq)
66{
67	trace_hfi1_txq_stop(txq);
68	if (atomic_inc_return(&txq->stops) == 1)
69		netif_stop_subqueue(txq->priv->netdev, txq->q_idx);
70}
71
72static void hfi1_ipoib_wake_txq(struct hfi1_ipoib_txq *txq)
73{
74	trace_hfi1_txq_wake(txq);
75	if (atomic_dec_and_test(&txq->stops))
76		netif_wake_subqueue(txq->priv->netdev, txq->q_idx);
77}
78
79static uint hfi1_ipoib_ring_hwat(struct hfi1_ipoib_txq *txq)
80{
81	return min_t(uint, txq->priv->netdev->tx_queue_len,
82		     txq->tx_ring.max_items - 1);
83}
84
85static uint hfi1_ipoib_ring_lwat(struct hfi1_ipoib_txq *txq)
86{
87	return min_t(uint, txq->priv->netdev->tx_queue_len,
88		     txq->tx_ring.max_items) >> 1;
89}
90
91static void hfi1_ipoib_check_queue_depth(struct hfi1_ipoib_txq *txq)
92{
93	++txq->sent_txreqs;
94	if (hfi1_ipoib_used(txq) >= hfi1_ipoib_ring_hwat(txq) &&
95	    !atomic_xchg(&txq->ring_full, 1)) {
96		trace_hfi1_txq_full(txq);
97		hfi1_ipoib_stop_txq(txq);
98	}
99}
100
101static void hfi1_ipoib_check_queue_stopped(struct hfi1_ipoib_txq *txq)
102{
103	struct net_device *dev = txq->priv->netdev;
104
105	/* If shutting down just return as queue state is irrelevant */
106	if (unlikely(dev->reg_state != NETREG_REGISTERED))
107		return;
108
109	/*
110	 * When the queue has been drained to less than half full it will be
111	 * restarted.
112	 * The size of the txreq ring is fixed at initialization.
113	 * The tx queue len can be adjusted upward while the interface is
114	 * running.
115	 * The tx queue len can be large enough to overflow the txreq_ring.
116	 * Use the minimum of the current tx_queue_len or the rings max txreqs
117	 * to protect against ring overflow.
118	 */
119	if (hfi1_ipoib_used(txq) < hfi1_ipoib_ring_lwat(txq) &&
120	    atomic_xchg(&txq->ring_full, 0)) {
121		trace_hfi1_txq_xmit_unstopped(txq);
122		hfi1_ipoib_wake_txq(txq);
123	}
124}
125
126static void hfi1_ipoib_free_tx(struct ipoib_txreq *tx, int budget)
127{
128	struct hfi1_ipoib_dev_priv *priv = tx->priv;
129
130	if (likely(!tx->sdma_status)) {
131		hfi1_ipoib_update_tx_netstats(priv, 1, tx->skb->len);
132	} else {
133		++priv->netdev->stats.tx_errors;
134		dd_dev_warn(priv->dd,
135			    "%s: Status = 0x%x pbc 0x%llx txq = %d sde = %d\n",
136			    __func__, tx->sdma_status,
137			    le64_to_cpu(tx->sdma_hdr.pbc), tx->txq->q_idx,
138			    tx->txq->sde->this_idx);
139	}
140
141	napi_consume_skb(tx->skb, budget);
142	sdma_txclean(priv->dd, &tx->txreq);
143	kmem_cache_free(priv->txreq_cache, tx);
144}
145
146static int hfi1_ipoib_drain_tx_ring(struct hfi1_ipoib_txq *txq, int budget)
147{
148	struct hfi1_ipoib_circ_buf *tx_ring = &txq->tx_ring;
149	unsigned long head;
150	unsigned long tail;
151	unsigned int max_tx;
152	int work_done;
153	int tx_count;
154
155	spin_lock_bh(&tx_ring->consumer_lock);
156
157	/* Read index before reading contents at that index. */
158	head = smp_load_acquire(&tx_ring->head);
159	tail = tx_ring->tail;
160	max_tx = tx_ring->max_items;
161
162	work_done = min_t(int, CIRC_CNT(head, tail, max_tx), budget);
163
164	for (tx_count = work_done; tx_count; tx_count--) {
165		hfi1_ipoib_free_tx(tx_ring->items[tail], budget);
166		tail = CIRC_NEXT(tail, max_tx);
167	}
168
169	atomic64_add(work_done, &txq->complete_txreqs);
170
171	/* Finished freeing tx items so store the tail value. */
172	smp_store_release(&tx_ring->tail, tail);
173
174	spin_unlock_bh(&tx_ring->consumer_lock);
175
176	hfi1_ipoib_check_queue_stopped(txq);
177
178	return work_done;
179}
180
181static int hfi1_ipoib_process_tx_ring(struct napi_struct *napi, int budget)
182{
183	struct hfi1_ipoib_dev_priv *priv = hfi1_ipoib_priv(napi->dev);
184	struct hfi1_ipoib_txq *txq = &priv->txqs[napi - priv->tx_napis];
185
186	int work_done = hfi1_ipoib_drain_tx_ring(txq, budget);
187
188	if (work_done < budget)
189		napi_complete_done(napi, work_done);
190
191	return work_done;
192}
193
194static void hfi1_ipoib_add_tx(struct ipoib_txreq *tx)
195{
196	struct hfi1_ipoib_circ_buf *tx_ring = &tx->txq->tx_ring;
197	unsigned long head;
198	unsigned long tail;
199	size_t max_tx;
200
201	spin_lock(&tx_ring->producer_lock);
202
203	head = tx_ring->head;
204	tail = READ_ONCE(tx_ring->tail);
205	max_tx = tx_ring->max_items;
206
207	if (likely(CIRC_SPACE(head, tail, max_tx))) {
208		tx_ring->items[head] = tx;
209
210		/* Finish storing txreq before incrementing head. */
211		smp_store_release(&tx_ring->head, CIRC_ADD(head, 1, max_tx));
212		napi_schedule(tx->txq->napi);
213	} else {
214		struct hfi1_ipoib_txq *txq = tx->txq;
215		struct hfi1_ipoib_dev_priv *priv = tx->priv;
216
217		/* Ring was full */
218		hfi1_ipoib_free_tx(tx, 0);
219		atomic64_inc(&txq->complete_txreqs);
220		dd_dev_dbg(priv->dd, "txq %d full.\n", txq->q_idx);
221	}
222
223	spin_unlock(&tx_ring->producer_lock);
224}
225
226static void hfi1_ipoib_sdma_complete(struct sdma_txreq *txreq, int status)
227{
228	struct ipoib_txreq *tx = container_of(txreq, struct ipoib_txreq, txreq);
229
230	tx->sdma_status = status;
231
232	hfi1_ipoib_add_tx(tx);
233}
234
235static int hfi1_ipoib_build_ulp_payload(struct ipoib_txreq *tx,
236					struct ipoib_txparms *txp)
237{
238	struct hfi1_devdata *dd = txp->dd;
239	struct sdma_txreq *txreq = &tx->txreq;
240	struct sk_buff *skb = tx->skb;
241	int ret = 0;
242	int i;
243
244	if (skb_headlen(skb)) {
245		ret = sdma_txadd_kvaddr(dd, txreq, skb->data, skb_headlen(skb));
246		if (unlikely(ret))
247			return ret;
248	}
249
250	for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
251		const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
252
253		ret = sdma_txadd_page(dd,
254				      txreq,
255				      skb_frag_page(frag),
256				      frag->bv_offset,
257				      skb_frag_size(frag),
258				      NULL, NULL, NULL);
259		if (unlikely(ret))
260			break;
261	}
262
263	return ret;
264}
265
266static int hfi1_ipoib_build_tx_desc(struct ipoib_txreq *tx,
267				    struct ipoib_txparms *txp)
268{
269	struct hfi1_devdata *dd = txp->dd;
270	struct sdma_txreq *txreq = &tx->txreq;
271	struct hfi1_sdma_header *sdma_hdr = &tx->sdma_hdr;
272	u16 pkt_bytes =
273		sizeof(sdma_hdr->pbc) + (txp->hdr_dwords << 2) + tx->skb->len;
274	int ret;
275
276	ret = sdma_txinit(txreq, 0, pkt_bytes, hfi1_ipoib_sdma_complete);
277	if (unlikely(ret))
278		return ret;
279
280	/* add pbc + headers */
281	ret = sdma_txadd_kvaddr(dd,
282				txreq,
283				sdma_hdr,
284				sizeof(sdma_hdr->pbc) + (txp->hdr_dwords << 2));
285	if (unlikely(ret))
286		return ret;
287
288	/* add the ulp payload */
289	return hfi1_ipoib_build_ulp_payload(tx, txp);
290}
291
292static void hfi1_ipoib_build_ib_tx_headers(struct ipoib_txreq *tx,
293					   struct ipoib_txparms *txp)
294{
295	struct hfi1_ipoib_dev_priv *priv = tx->priv;
296	struct hfi1_sdma_header *sdma_hdr = &tx->sdma_hdr;
297	struct sk_buff *skb = tx->skb;
298	struct hfi1_pportdata *ppd = ppd_from_ibp(txp->ibp);
299	struct rdma_ah_attr *ah_attr = txp->ah_attr;
300	struct ib_other_headers *ohdr;
301	struct ib_grh *grh;
302	u16 dwords;
303	u16 slid;
304	u16 dlid;
305	u16 lrh0;
306	u32 bth0;
307	u32 sqpn = (u32)(priv->netdev->dev_addr[1] << 16 |
308			 priv->netdev->dev_addr[2] << 8 |
309			 priv->netdev->dev_addr[3]);
310	u16 payload_dwords;
311	u8 pad_cnt;
312
313	pad_cnt = -skb->len & 3;
314
315	/* Includes ICRC */
316	payload_dwords = ((skb->len + pad_cnt) >> 2) + SIZE_OF_CRC;
317
318	/* header size in dwords LRH+BTH+DETH = (8+12+8)/4. */
319	txp->hdr_dwords = 7;
320
321	if (rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH) {
322		grh = &sdma_hdr->hdr.ibh.u.l.grh;
323		txp->hdr_dwords +=
324			hfi1_make_grh(txp->ibp,
325				      grh,
326				      rdma_ah_read_grh(ah_attr),
327				      txp->hdr_dwords - LRH_9B_DWORDS,
328				      payload_dwords);
329		lrh0 = HFI1_LRH_GRH;
330		ohdr = &sdma_hdr->hdr.ibh.u.l.oth;
331	} else {
332		lrh0 = HFI1_LRH_BTH;
333		ohdr = &sdma_hdr->hdr.ibh.u.oth;
334	}
335
336	lrh0 |= (rdma_ah_get_sl(ah_attr) & 0xf) << 4;
337	lrh0 |= (txp->flow.sc5 & 0xf) << 12;
338
339	dlid = opa_get_lid(rdma_ah_get_dlid(ah_attr), 9B);
340	if (dlid == be16_to_cpu(IB_LID_PERMISSIVE)) {
341		slid = be16_to_cpu(IB_LID_PERMISSIVE);
342	} else {
343		u16 lid = (u16)ppd->lid;
344
345		if (lid) {
346			lid |= rdma_ah_get_path_bits(ah_attr) &
347				((1 << ppd->lmc) - 1);
348			slid = lid;
349		} else {
350			slid = be16_to_cpu(IB_LID_PERMISSIVE);
351		}
352	}
353
354	/* Includes ICRC */
355	dwords = txp->hdr_dwords + payload_dwords;
356
357	/* Build the lrh */
358	sdma_hdr->hdr.hdr_type = HFI1_PKT_TYPE_9B;
359	hfi1_make_ib_hdr(&sdma_hdr->hdr.ibh, lrh0, dwords, dlid, slid);
360
361	/* Build the bth */
362	bth0 = (IB_OPCODE_UD_SEND_ONLY << 24) | (pad_cnt << 20) | priv->pkey;
363
364	ohdr->bth[0] = cpu_to_be32(bth0);
365	ohdr->bth[1] = cpu_to_be32(txp->dqpn);
366	ohdr->bth[2] = cpu_to_be32(mask_psn((u32)txp->txq->sent_txreqs));
367
368	/* Build the deth */
369	ohdr->u.ud.deth[0] = cpu_to_be32(priv->qkey);
370	ohdr->u.ud.deth[1] = cpu_to_be32((txp->entropy <<
371					  HFI1_IPOIB_ENTROPY_SHIFT) | sqpn);
372
373	/* Construct the pbc. */
374	sdma_hdr->pbc =
375		cpu_to_le64(create_pbc(ppd,
376				       ib_is_sc5(txp->flow.sc5) <<
377							      PBC_DC_INFO_SHIFT,
378				       0,
379				       sc_to_vlt(priv->dd, txp->flow.sc5),
380				       dwords - SIZE_OF_CRC +
381						(sizeof(sdma_hdr->pbc) >> 2)));
382}
383
384static struct ipoib_txreq *hfi1_ipoib_send_dma_common(struct net_device *dev,
385						      struct sk_buff *skb,
386						      struct ipoib_txparms *txp)
387{
388	struct hfi1_ipoib_dev_priv *priv = hfi1_ipoib_priv(dev);
389	struct ipoib_txreq *tx;
390	int ret;
391
392	tx = kmem_cache_alloc_node(priv->txreq_cache,
393				   GFP_ATOMIC,
394				   priv->dd->node);
395	if (unlikely(!tx))
396		return ERR_PTR(-ENOMEM);
397
398	/* so that we can test if the sdma descriptors are there */
399	tx->txreq.num_desc = 0;
400	tx->priv = priv;
401	tx->txq = txp->txq;
402	tx->skb = skb;
403	INIT_LIST_HEAD(&tx->txreq.list);
404
405	hfi1_ipoib_build_ib_tx_headers(tx, txp);
406
407	ret = hfi1_ipoib_build_tx_desc(tx, txp);
408	if (likely(!ret)) {
409		if (txp->txq->flow.as_int != txp->flow.as_int) {
410			txp->txq->flow.tx_queue = txp->flow.tx_queue;
411			txp->txq->flow.sc5 = txp->flow.sc5;
412			txp->txq->sde =
413				sdma_select_engine_sc(priv->dd,
414						      txp->flow.tx_queue,
415						      txp->flow.sc5);
416			trace_hfi1_flow_switch(txp->txq);
417		}
418
419		return tx;
420	}
421
422	sdma_txclean(priv->dd, &tx->txreq);
423	kmem_cache_free(priv->txreq_cache, tx);
424
425	return ERR_PTR(ret);
426}
427
428static int hfi1_ipoib_submit_tx_list(struct net_device *dev,
429				     struct hfi1_ipoib_txq *txq)
430{
431	int ret;
432	u16 count_out;
433
434	ret = sdma_send_txlist(txq->sde,
435			       iowait_get_ib_work(&txq->wait),
436			       &txq->tx_list,
437			       &count_out);
438	if (likely(!ret) || ret == -EBUSY || ret == -ECOMM)
439		return ret;
440
441	dd_dev_warn(txq->priv->dd, "cannot send skb tx list, err %d.\n", ret);
442
443	return ret;
444}
445
446static int hfi1_ipoib_flush_tx_list(struct net_device *dev,
447				    struct hfi1_ipoib_txq *txq)
448{
449	int ret = 0;
450
451	if (!list_empty(&txq->tx_list)) {
452		/* Flush the current list */
453		ret = hfi1_ipoib_submit_tx_list(dev, txq);
454
455		if (unlikely(ret))
456			if (ret != -EBUSY)
457				++dev->stats.tx_carrier_errors;
458	}
459
460	return ret;
461}
462
463static int hfi1_ipoib_submit_tx(struct hfi1_ipoib_txq *txq,
464				struct ipoib_txreq *tx)
465{
466	int ret;
467
468	ret = sdma_send_txreq(txq->sde,
469			      iowait_get_ib_work(&txq->wait),
470			      &tx->txreq,
471			      txq->pkts_sent);
472	if (likely(!ret)) {
473		txq->pkts_sent = true;
474		iowait_starve_clear(txq->pkts_sent, &txq->wait);
475	}
476
477	return ret;
478}
479
480static int hfi1_ipoib_send_dma_single(struct net_device *dev,
481				      struct sk_buff *skb,
482				      struct ipoib_txparms *txp)
483{
484	struct hfi1_ipoib_dev_priv *priv = hfi1_ipoib_priv(dev);
485	struct hfi1_ipoib_txq *txq = txp->txq;
486	struct ipoib_txreq *tx;
487	int ret;
488
489	tx = hfi1_ipoib_send_dma_common(dev, skb, txp);
490	if (IS_ERR(tx)) {
491		int ret = PTR_ERR(tx);
492
493		dev_kfree_skb_any(skb);
494
495		if (ret == -ENOMEM)
496			++dev->stats.tx_errors;
497		else
498			++dev->stats.tx_carrier_errors;
499
500		return NETDEV_TX_OK;
501	}
502
503	ret = hfi1_ipoib_submit_tx(txq, tx);
504	if (likely(!ret)) {
505tx_ok:
506		trace_sdma_output_ibhdr(tx->priv->dd,
507					&tx->sdma_hdr.hdr,
508					ib_is_sc5(txp->flow.sc5));
509		hfi1_ipoib_check_queue_depth(txq);
510		return NETDEV_TX_OK;
511	}
512
513	txq->pkts_sent = false;
514
515	if (ret == -EBUSY || ret == -ECOMM)
516		goto tx_ok;
517
518	sdma_txclean(priv->dd, &tx->txreq);
519	dev_kfree_skb_any(skb);
520	kmem_cache_free(priv->txreq_cache, tx);
521	++dev->stats.tx_carrier_errors;
522
523	return NETDEV_TX_OK;
524}
525
526static int hfi1_ipoib_send_dma_list(struct net_device *dev,
527				    struct sk_buff *skb,
528				    struct ipoib_txparms *txp)
529{
530	struct hfi1_ipoib_txq *txq = txp->txq;
531	struct ipoib_txreq *tx;
532
533	/* Has the flow change ? */
534	if (txq->flow.as_int != txp->flow.as_int) {
535		int ret;
536
537		trace_hfi1_flow_flush(txq);
538		ret = hfi1_ipoib_flush_tx_list(dev, txq);
539		if (unlikely(ret)) {
540			if (ret == -EBUSY)
541				++dev->stats.tx_dropped;
542			dev_kfree_skb_any(skb);
543			return NETDEV_TX_OK;
544		}
545	}
546	tx = hfi1_ipoib_send_dma_common(dev, skb, txp);
547	if (IS_ERR(tx)) {
548		int ret = PTR_ERR(tx);
549
550		dev_kfree_skb_any(skb);
551
552		if (ret == -ENOMEM)
553			++dev->stats.tx_errors;
554		else
555			++dev->stats.tx_carrier_errors;
556
557		return NETDEV_TX_OK;
558	}
559
560	list_add_tail(&tx->txreq.list, &txq->tx_list);
561
562	hfi1_ipoib_check_queue_depth(txq);
563
564	trace_sdma_output_ibhdr(tx->priv->dd,
565				&tx->sdma_hdr.hdr,
566				ib_is_sc5(txp->flow.sc5));
567
568	if (!netdev_xmit_more())
569		(void)hfi1_ipoib_flush_tx_list(dev, txq);
570
571	return NETDEV_TX_OK;
572}
573
574static u8 hfi1_ipoib_calc_entropy(struct sk_buff *skb)
575{
576	if (skb_transport_header_was_set(skb)) {
577		u8 *hdr = (u8 *)skb_transport_header(skb);
578
579		return (hdr[0] ^ hdr[1] ^ hdr[2] ^ hdr[3]);
580	}
581
582	return (u8)skb_get_queue_mapping(skb);
583}
584
585int hfi1_ipoib_send_dma(struct net_device *dev,
586			struct sk_buff *skb,
587			struct ib_ah *address,
588			u32 dqpn)
589{
590	struct hfi1_ipoib_dev_priv *priv = hfi1_ipoib_priv(dev);
591	struct ipoib_txparms txp;
592	struct rdma_netdev *rn = netdev_priv(dev);
593
594	if (unlikely(skb->len > rn->mtu + HFI1_IPOIB_ENCAP_LEN)) {
595		dd_dev_warn(priv->dd, "packet len %d (> %d) too long to send, dropping\n",
596			    skb->len,
597			    rn->mtu + HFI1_IPOIB_ENCAP_LEN);
598		++dev->stats.tx_dropped;
599		++dev->stats.tx_errors;
600		dev_kfree_skb_any(skb);
601		return NETDEV_TX_OK;
602	}
603
604	txp.dd = priv->dd;
605	txp.ah_attr = &ibah_to_rvtah(address)->attr;
606	txp.ibp = to_iport(priv->device, priv->port_num);
607	txp.txq = &priv->txqs[skb_get_queue_mapping(skb)];
608	txp.dqpn = dqpn;
609	txp.flow.sc5 = txp.ibp->sl_to_sc[rdma_ah_get_sl(txp.ah_attr)];
610	txp.flow.tx_queue = (u8)skb_get_queue_mapping(skb);
611	txp.entropy = hfi1_ipoib_calc_entropy(skb);
612
613	if (netdev_xmit_more() || !list_empty(&txp.txq->tx_list))
614		return hfi1_ipoib_send_dma_list(dev, skb, &txp);
615
616	return hfi1_ipoib_send_dma_single(dev, skb,  &txp);
617}
618
619/*
620 * hfi1_ipoib_sdma_sleep - ipoib sdma sleep function
621 *
622 * This function gets called from sdma_send_txreq() when there are not enough
623 * sdma descriptors available to send the packet. It adds Tx queue's wait
624 * structure to sdma engine's dmawait list to be woken up when descriptors
625 * become available.
626 */
627static int hfi1_ipoib_sdma_sleep(struct sdma_engine *sde,
628				 struct iowait_work *wait,
629				 struct sdma_txreq *txreq,
630				 uint seq,
631				 bool pkts_sent)
632{
633	struct hfi1_ipoib_txq *txq =
634		container_of(wait->iow, struct hfi1_ipoib_txq, wait);
635
636	write_seqlock(&sde->waitlock);
637
638	if (likely(txq->priv->netdev->reg_state == NETREG_REGISTERED)) {
639		if (sdma_progress(sde, seq, txreq)) {
640			write_sequnlock(&sde->waitlock);
641			return -EAGAIN;
642		}
643
644		if (list_empty(&txreq->list))
645			/* came from non-list submit */
646			list_add_tail(&txreq->list, &txq->tx_list);
647		if (list_empty(&txq->wait.list)) {
648			if (!atomic_xchg(&txq->no_desc, 1)) {
649				trace_hfi1_txq_queued(txq);
650				hfi1_ipoib_stop_txq(txq);
651			}
652			iowait_queue(pkts_sent, wait->iow, &sde->dmawait);
653		}
654
655		write_sequnlock(&sde->waitlock);
656		return -EBUSY;
657	}
658
659	write_sequnlock(&sde->waitlock);
660	return -EINVAL;
661}
662
663/*
664 * hfi1_ipoib_sdma_wakeup - ipoib sdma wakeup function
665 *
666 * This function gets called when SDMA descriptors becomes available and Tx
667 * queue's wait structure was previously added to sdma engine's dmawait list.
668 */
669static void hfi1_ipoib_sdma_wakeup(struct iowait *wait, int reason)
670{
671	struct hfi1_ipoib_txq *txq =
672		container_of(wait, struct hfi1_ipoib_txq, wait);
673
674	trace_hfi1_txq_wakeup(txq);
675	if (likely(txq->priv->netdev->reg_state == NETREG_REGISTERED))
676		iowait_schedule(wait, system_highpri_wq, WORK_CPU_UNBOUND);
677}
678
679static void hfi1_ipoib_flush_txq(struct work_struct *work)
680{
681	struct iowait_work *ioww =
682		container_of(work, struct iowait_work, iowork);
683	struct iowait *wait = iowait_ioww_to_iow(ioww);
684	struct hfi1_ipoib_txq *txq =
685		container_of(wait, struct hfi1_ipoib_txq, wait);
686	struct net_device *dev = txq->priv->netdev;
687
688	if (likely(dev->reg_state == NETREG_REGISTERED) &&
689	    likely(!hfi1_ipoib_flush_tx_list(dev, txq)))
690		if (atomic_xchg(&txq->no_desc, 0))
691			hfi1_ipoib_wake_txq(txq);
692}
693
694int hfi1_ipoib_txreq_init(struct hfi1_ipoib_dev_priv *priv)
695{
696	struct net_device *dev = priv->netdev;
697	char buf[HFI1_IPOIB_TXREQ_NAME_LEN];
698	unsigned long tx_ring_size;
699	int i;
700
701	/*
702	 * Ring holds 1 less than tx_ring_size
703	 * Round up to next power of 2 in order to hold at least tx_queue_len
704	 */
705	tx_ring_size = roundup_pow_of_two((unsigned long)dev->tx_queue_len + 1);
706
707	snprintf(buf, sizeof(buf), "hfi1_%u_ipoib_txreq_cache", priv->dd->unit);
708	priv->txreq_cache = kmem_cache_create(buf,
709					      sizeof(struct ipoib_txreq),
710					      0,
711					      0,
712					      NULL);
713	if (!priv->txreq_cache)
714		return -ENOMEM;
715
716	priv->tx_napis = kcalloc_node(dev->num_tx_queues,
717				      sizeof(struct napi_struct),
718				      GFP_KERNEL,
719				      priv->dd->node);
720	if (!priv->tx_napis)
721		goto free_txreq_cache;
722
723	priv->txqs = kcalloc_node(dev->num_tx_queues,
724				  sizeof(struct hfi1_ipoib_txq),
725				  GFP_KERNEL,
726				  priv->dd->node);
727	if (!priv->txqs)
728		goto free_tx_napis;
729
730	for (i = 0; i < dev->num_tx_queues; i++) {
731		struct hfi1_ipoib_txq *txq = &priv->txqs[i];
732
733		iowait_init(&txq->wait,
734			    0,
735			    hfi1_ipoib_flush_txq,
736			    NULL,
737			    hfi1_ipoib_sdma_sleep,
738			    hfi1_ipoib_sdma_wakeup,
739			    NULL,
740			    NULL);
741		txq->priv = priv;
742		txq->sde = NULL;
743		INIT_LIST_HEAD(&txq->tx_list);
744		atomic64_set(&txq->complete_txreqs, 0);
745		atomic_set(&txq->stops, 0);
746		atomic_set(&txq->ring_full, 0);
747		atomic_set(&txq->no_desc, 0);
748		txq->q_idx = i;
749		txq->flow.tx_queue = 0xff;
750		txq->flow.sc5 = 0xff;
751		txq->pkts_sent = false;
752
753		netdev_queue_numa_node_write(netdev_get_tx_queue(dev, i),
754					     priv->dd->node);
755
756		txq->tx_ring.items =
757			kcalloc_node(tx_ring_size,
758				     sizeof(struct ipoib_txreq *),
759				     GFP_KERNEL, priv->dd->node);
760		if (!txq->tx_ring.items)
761			goto free_txqs;
762
763		spin_lock_init(&txq->tx_ring.producer_lock);
764		spin_lock_init(&txq->tx_ring.consumer_lock);
765		txq->tx_ring.max_items = tx_ring_size;
766
767		txq->napi = &priv->tx_napis[i];
768		netif_tx_napi_add(dev, txq->napi,
769				  hfi1_ipoib_process_tx_ring,
770				  NAPI_POLL_WEIGHT);
771	}
772
773	return 0;
774
775free_txqs:
776	for (i--; i >= 0; i--) {
777		struct hfi1_ipoib_txq *txq = &priv->txqs[i];
778
779		netif_napi_del(txq->napi);
780		kfree(txq->tx_ring.items);
781	}
782
783	kfree(priv->txqs);
784	priv->txqs = NULL;
785
786free_tx_napis:
787	kfree(priv->tx_napis);
788	priv->tx_napis = NULL;
789
790free_txreq_cache:
791	kmem_cache_destroy(priv->txreq_cache);
792	priv->txreq_cache = NULL;
793	return -ENOMEM;
794}
795
796static void hfi1_ipoib_drain_tx_list(struct hfi1_ipoib_txq *txq)
797{
798	struct sdma_txreq *txreq;
799	struct sdma_txreq *txreq_tmp;
800	atomic64_t *complete_txreqs = &txq->complete_txreqs;
801
802	list_for_each_entry_safe(txreq, txreq_tmp, &txq->tx_list, list) {
803		struct ipoib_txreq *tx =
804			container_of(txreq, struct ipoib_txreq, txreq);
805
806		list_del(&txreq->list);
807		sdma_txclean(txq->priv->dd, &tx->txreq);
808		dev_kfree_skb_any(tx->skb);
809		kmem_cache_free(txq->priv->txreq_cache, tx);
810		atomic64_inc(complete_txreqs);
811	}
812
813	if (hfi1_ipoib_used(txq))
814		dd_dev_warn(txq->priv->dd,
815			    "txq %d not empty found %llu requests\n",
816			    txq->q_idx,
817			    hfi1_ipoib_txreqs(txq->sent_txreqs,
818					      atomic64_read(complete_txreqs)));
819}
820
821void hfi1_ipoib_txreq_deinit(struct hfi1_ipoib_dev_priv *priv)
822{
823	int i;
824
825	for (i = 0; i < priv->netdev->num_tx_queues; i++) {
826		struct hfi1_ipoib_txq *txq = &priv->txqs[i];
827
828		iowait_cancel_work(&txq->wait);
829		iowait_sdma_drain(&txq->wait);
830		hfi1_ipoib_drain_tx_list(txq);
831		netif_napi_del(txq->napi);
832		(void)hfi1_ipoib_drain_tx_ring(txq, txq->tx_ring.max_items);
833		kfree(txq->tx_ring.items);
834	}
835
836	kfree(priv->txqs);
837	priv->txqs = NULL;
838
839	kfree(priv->tx_napis);
840	priv->tx_napis = NULL;
841
842	kmem_cache_destroy(priv->txreq_cache);
843	priv->txreq_cache = NULL;
844}
845
846void hfi1_ipoib_napi_tx_enable(struct net_device *dev)
847{
848	struct hfi1_ipoib_dev_priv *priv = hfi1_ipoib_priv(dev);
849	int i;
850
851	for (i = 0; i < dev->num_tx_queues; i++) {
852		struct hfi1_ipoib_txq *txq = &priv->txqs[i];
853
854		napi_enable(txq->napi);
855	}
856}
857
858void hfi1_ipoib_napi_tx_disable(struct net_device *dev)
859{
860	struct hfi1_ipoib_dev_priv *priv = hfi1_ipoib_priv(dev);
861	int i;
862
863	for (i = 0; i < dev->num_tx_queues; i++) {
864		struct hfi1_ipoib_txq *txq = &priv->txqs[i];
865
866		napi_disable(txq->napi);
867		(void)hfi1_ipoib_drain_tx_ring(txq, txq->tx_ring.max_items);
868	}
869}
870